Introducing DR34MSC4PE's first test training of a Hunyuan LoRA: Reverse Cowgirl.
Based on the guide found here: https://civarchive.com/articles/9798/training-a-lora-for-hunyuan-video-on-windows
===========================================================================
LoRA trained with 48 short clips at 2 seconds each of this specific position. Trained on 2x3090s, took about 24 hours.
Recommended Settings
These resolutions seem to work pretty well:
640x480
960x720
720X544
544x960
Steps: 20-40
CFG: 4-6
euler/normal or simple
ipndm/beta
deis/beta
dpm++2m/normal or beta
Experiment with your own settings to see what works.
Sample Prompts:
c0wg1rl, full body, wide angle, a cute woman with dark hair riding a penis in reverse cowgirl position, She has large supple breasts and is wearing only a short blue skirt. Her facial expression is one of pleasure with her eyes closed and mouth open as she moans. A man's penis is thrusting in and out of her vagina. The scene takes place in a dark alley on the cracked and icy pavement illuminated by streetlamp
Workflow.json is in the Dataset download zip.
===========================================================================
DR34MSC4PE is
c0ur4ge (trainer/qa/inference code) /
ERA5ER (captions/tooling code/data/qa)
Check out our Flux model, C4PACITOR!
Like our work? Buy us a coffee: https://ko-fi.com/dr34msc4pe
Please generate responsibly!
Description
Trained on the bf16 model, have not tested with fp8 or GGUF versions.
FAQ
Comments (14)
Thank you! I don't know if it's intentional but downloading the training images form the link in the right panel gives me a zip file with a workflow which is also great, but people may except to find the training data there.
Thanks!
Yes, that just has the workflow. I wasn't able to edit the name of it.
If there's enough interest, I'll add in the txt files and config files.
EDIT: Added the txt and training files.
Works good, but with other prompt. use the prompt of the other Reverse Cowgirl lora and add the keyword "c0wg1rl,"
Thank you! I'm not sure yet what prompts work best for Hunyuan, so I encourage everyone to share their prompts and workflows if they can!
This works very well and a good variety. Feels very natural! Hope you make more Loras with that level of quality =)
Thanks, I appreciate it. I definitely have more in the diffusion pipeline, stay tuned!
Your lora is great! , can I have a footjob with lora, thank you very much!
Requests are kinda in bad taste, although this guys actually worth asking it seems. He does have a support link, that would likely be the appropriate way to specify exact requests.
Very powerful lora, 9/10 generations are perfect.
So will LORAs like these work with the new image-2-video model?
i used it with i2V workflows, and it works perfect every time.
Hey, i'd be down to train a wan version of this if you'd be open to (privately) sharing the dataset. Would credit you ofc.
Hey thanks for the offer! I actually have a Wan version in the oven right now. Hoping to share in the near future.
Sweet, looking forward to it. LMK if you want to combine with the data I used for my frontal lora (https://civitai.com/models/1321853/frontal-sex), I have the RC separated out already I think.